Biometric templates attached to agent memory separate owner from non-owner queries by up to 27 points F1 in shared deployments

arXiv·low signalXBlueskyLinkedInCopy link

Two companion papers address the access-control question semantic retrieval cannot answer in multi-user agent deployments: whether a memory block belongs to the person asking right now. Bio-Memory augments each A-Mem atomic note with a biometric embedding and forms the retrieval candidate pool by biometric match before semantic ranking, and on LoCoMo in a 10-user shared-agent setting across 7 face and 10 palmprint protocols it consistently separates owner from non-owner queries, with the largest average gap reaching 27.29%/21.15% F1/BLEU-1 on CALFW. Bio-MemArt applies the same idea one layer down, attaching normalized biometric templates to KV-cache memory blocks so reusable cache blocks are filtered by the current user's probe before MemArt retrieval and cache reuse.